UP Baguio Seminar, 02/02/ 2007 1.Global optimization problems in various industrial fields 1.1 Car shape optimization in automotive industry (in collaboration with PSA Peugeot Citroën) 1.2 Optical fiber optimization in telecommunication industry (in collaboration with Alcatel) 2. Existing optimization methods 3. Presentation and validation of a new hybrid method 4. Application to Laurent Dumas Laboratoire Jacques-Louis Lions, UPMC, Paris, [email protected]Hybrid optimization and application to various industrial problems
27
Embed
UP Baguio Seminar, 02/02/2007 1.Global optimization problems in various industrial fieldsGlobal optimization problems in various industrial fields 1.1.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
UP Baguio Seminar, 02/02/2007
1. Global optimization problems in various industrial fields1.1 Car shape optimization in automotive industry
(in collaboration with PSA Peugeot Citroën)
1.2 Optical fiber optimization in telecommunication industry
(in collaboration with Alcatel)
2. Existing optimization methods
3. Presentation and validation of a new hybrid method
4. Application to the previous industrial problems
Hybrid optimization and application to various industrial problems
UP Baguio Seminar, 02/02/2007
1.1.1 1.1.1 Car shapeCar shape optimization for fuel consumption reduction optimization for fuel consumption reduction
(joint work with PSA Peugeot Citroën)(joint work with PSA Peugeot Citroën)
Possibility of action to reduce consumption :Possibility of action to reduce consumption : - Motors evolution
- Weight reduction
- Car shape optimization
Aerodynamic drag
74%
Others
26%
Fuel consumption repartition for a car at 120 km/h:Fuel consumption repartition for a car at 120 km/h:
UP Baguio Seminar, 02/02/2007
Fx
Fz
•Aerodynamic drag force:FFF P
S
SSS
P
dSvCfVF
dSPdSPdSCpVF
//
2
2
2
1
2
1
• Drag coefficient : Cx1
SCxVFx 2
2
1.1.2 Definition of the drag coefficient1.1.2 Definition of the drag coefficient
UP Baguio Seminar, 02/02/2007
Lower exterior shape
CoolingWheels15%
Upper exterior shape
40%25 to 30 %
10 to 15%
Others5%
65%to 70 % of Cx dépends on the exterior shape
Aerodynamic flow at the rear of Peugeot 206 (DRIA)
1.1.3 Origins of the drag coefficient1.1.3 Origins of the drag coefficient
UP Baguio Seminar, 02/02/2007
fastSolving
constrained problems
robust
Parametrization of the geometry and definition of
a cost function
1.1.4 General formulation of the optimization problem1.1.4 General formulation of the optimization problem
Détermination of N control parameters
Optimisation method of J : RN R
UP Baguio Seminar, 02/02/2007
1.2.1 Optical fiber optimization for the construction of a mono/multi channel wavelength filter
> Such filters can be obtained by using an optical fiber called FBG (Fiber Bragg Grating) having a fast periodic modulation of its refractive index in the core:
> The index variation can be optimized in order to give the desired reflectivity spectrum: inverse problem
m
(reflectivity spectrum)
(joint work with Alcatel)(joint work with Alcatel)
UP Baguio Seminar, 02/02/2007
1.2.2 Mathematical modelization of a FBG
• The refractive index of a FBG is expressed through a quasi-sinusoïdal function in the longitudinal direction z:
n(z)=n0+n(z) cos(2z/0) z [0, L]
with the following notations:
n0 : index refraction of the core
0: nominal period of the FBG
n(z): slowly varying amplitude (also called apodisation)
• The inverse-type optimization problem will consist in finding the ‘best’ apodisation function leading to the desired reflectivity spectrum.
UP Baguio Seminar, 02/02/2007
1.2.3 Computation of the reflectivity spectrum of a FBG
•The reflectivity spectrum is a function R() =| r() |2 where
r() = bB(0,) / bF(0,)
• In the above expression, the enveloppes of the forward and backward propagating waves are obtained by the resolution of the following system of coupled ODE’s:
FBG with Gaussian apodisation FBG with raised-cosine apodisation
FBG with weak constant apodisation ( n=1E-4) FBG with strong constant apodistion (n=4E-4)
•Four examples of reflectivity spectra are displayed below corresponding to four different FBG (L=20cm, n0=1.45, B=1550nm):
UP Baguio Seminar, 02/02/2007
fastSolving
constrained problems
robust
Parametrization of the geometry and definition of
a cost function
1.2.5 General formulation of the optimization problem1.2.5 General formulation of the optimization problem
Détermination of N control parameters
Optimisation method of J : RN R
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07-0.5
0
0.5
1
1.5
2
2.5
3
3.5
4x 10
-4
z
dn(z
)
enveloppe de l'indice
dn(premiere generation)dn(derniere generation)
UP Baguio Seminar, 02/02/2007
2. Existing optimization methods
• Numerous global or local optimization methods exist to find the minimum of a given function J : O RN R . Among them, the main two classes are the following:
Deterministic gradient-based methods (steepest descent, quasi Newton, etc…) where the gradient of the cost function is needed.
Stochastic methods (simulated annealing, genetic algorithms, evolution strategies, … ) seeking for a global optimum.
•Both methods will be applied and compared on the classical Rastrigin function with n parameters which exhibits many local minima but only one global:
Rastrigin function with 2 parameters
nx2cosxXFn
1i
i2iRastrigin
UP Baguio Seminar, 02/02/2007
2.1 Principle of gradient-based methods
• Gradient based methods use the gradient of the cost function J in order to construct a sequence of points (xk), with decreasing values by J.
• At each iteration, once a descent direction is chosen, a linesearch principle is used to ensure a sufficient decrease in this direction.
• In a gradient-based method such as the steepest descent or the BFGS method, the main difficulty is to compute the gradient of the cost function,x J(x). In some
cases, the gradient is even not achievable and can only be approximated by finite differences.
UP Baguio Seminar, 02/02/2007
Initialisation: random selection of a population of Np ‘individuals’ associated to different values of the parameters xO Rn.
Evaluation: to each individual is associated a ‘fitness’ value inversely proportional to the cost function J to minimize.
The population is evolving at each generation through three ‘Darwinian’ principles of selection, crossover and mutation (detailed below).
After Ng generations, the average and the best fitness value of individuals have improved.
•Selection: the individuals are selected through a non-uniform random wheel:
•Crossover: (with probability pc): starting from two individuals x and y (, two new individuals are randomly generated from a barycentric combination of each component of x and y.
•Mutation: (with probability pm): starting from an individual x, a new individual can be randomly created with a normal law centered in x
• In general, a one-elitism strategy is included in order to keep the current best individual at the next generation.
UP Baguio Seminar, 02/02/2007
Initialisation: same as in GA (Np individuals randomly chosen)
The Darwinian principles consist only of crossover and mutation.
The number of generated offspring is higher than Np. The next generation of Np individuals is then obtained after a deterministic selection among parents and offsprings (called plus selection) or among offsprings (comma selection).
2.4 Principle of evolution strategies (ES)
Initialisation
Darwinian principles :Crossover- Mutation
Convergencetest
Ngen=Ngen+1
No
Evaluation of the population
Deterministic selection
UP Baguio Seminar, 02/02/2007
Convergence speed
Global minimization No regularity conditions on J Robustness Multi-objective Parallélisable
Local minimization Gradient evaluation Not multi-objective
Convergence speed
drawbacksadvantagesdrawbacksadvantages
Stochastic methods (GA, ES)Gradient-based methods
2.5 Comparison of the two types of methods
UP Baguio Seminar, 02/02/2007
3.1 Description of hybrid methods
Example of convergence history
Cost function evaluation number
Co
st f
un
ctio
n
gradient
• Idea : couple a stochastic optimization method (GA or ES) with a deterministic one in order to improve its convergence speed
•Principle: application of a descent type method to one or a few well chosen individuals at well chosen moments
GA or ES
UP Baguio Seminar, 02/02/2007
3.2 Description of the coupling principle
Final local search
END
BEGIN
G2L
Stopping criterion
Global search (AG) Local search (gradient)
Stopping criterion
L2G
• The coupling principle between the global search and the local search is done on an adaptative way with respect to two coefficients G2L and L2G.
UP Baguio Seminar, 02/02/2007
O
O
O
• elements of the population O center of mass of clusters
3.3 Choice of the elements for local search
• A clustering method is used. It consists to divide the population into N regularly distributed sub-population. The local search process is then applied to every best element of each cluster.
UP Baguio Seminar, 02/02/2007
3.4 Another way to improve GA: approximate evaluations
4.5 Application 2: design of a monochannel FBG filter
(L.D, O. Durand, B. Ivorra, B. Mohammadi, IJCSE, 2006)
UP Baguio Seminar, 02/02/2007
5. Conclusions
• Various global optimization methods have been developped, all consisting in a convergence acceleration of stochastic methods (GA or ES) by incorporating a new ‘intelligent’ mutation principle (namely a gradient-based method). The observed acceleration convergence speed ranges from a factor 2 to 10.
• Different industrial problems have been solved more efficiently with the help of these new methods, either direct (drag minimization, etc…) or inverse problems (filter design).
… presentation and Scilab scripts available next week at: